Skip to content

CHANGELOG.md

Latest commit

 

History

History
483 lines (339 loc) · 20.6 KB

CHANGELOG.md

File metadata and controls

483 lines (339 loc) · 20.6 KB
Raw

Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

[Unreleased]

[0.10.10] - 2026-07-03

Security

  • quinn-proto 0.11.14 → 0.11.15 (RUSTSEC-2026-0185) — fixes remote memory exhaustion via unbounded out-of-order stream reassembly. CVSS 7.5 (high). (#132)

Added

  • dakera_tif_evaluate tool (Power tier) — wraps the T-I-F (Truth-Indeterminacy-Falsity) reliability workflow into a single MCP call. Given a memory_id (and optional agent_id), it fetches the memory's feedback history, counts upvote/downvote/flag signals, computes normalised T-I-F scores, and returns a reuse classification (confident_reuse, surface_contradiction, ask_clarification, verify_before_use). Pure arithmetic on feedback signals — no external LLM. Brings the all profile to 87 tools and power to 69.

Fixed

  • Memory feedback tools missing agent_id — dakera_memory_feedback and dakera_memory_feedback_get now forward the required agent_id field; all agents' feedback calls previously returned 422. Restores the org-wide memory reinforcement loop. (#136)
  • CI integration tests — added explicit GHCR docker/login-action step before pulling the dakera image; stale credentials on self-hosted ARM runners caused denied: denied on the public image pull.
  • Docker port mapping — corrected 3300:3300 → 3300:3000 in documentation. (#133)
  • Playground badge URL — badge now links to dakera.ai/playground (frontend) instead of the API endpoint. (#131)
  • MCP Registry org case — restored Dakera-AI casing in repository.url after a previous normalisation incorrectly lowercased it. (DAK-7032, #128)

CI & Distribution

  • ARM runner lockfile-aware stale cache purge to prevent spurious compilation failures. (#135)
  • Bumped actions/checkout v4 → v7 and actions/cache v5 → v6. (#127, #130, #134)

[0.10.4] - 2026-05-18

Added

  • Glama.ai integration files — added glama.json claim file and Glama score badge to README for server directory discoverability. (#99, #100)
  • Improved tool descriptions for TDQS scoring — rewrote all 86 MCP tool descriptions with richer context, parameter explanations, and usage examples to improve Glama quality scores. (#102)
  • GitHub Release binary assets — release workflow now uploads pre-built linux x86_64 binary, tarball, and SHA256 checksums to GitHub Releases. (#101)

[0.10.3] - 2026-05-18

Added

  • Crate metadata for Glama.ai discoverability — added authors, homepage, documentation, readme, keywords, and categories to Cargo.toml. Fixes "No maintainers listed" on Glama.ai registry. (#98)

[0.10.2] - 2026-05-18

Fixed

  • MCP pagination page size increased from 20 to 100 — all profiles (core=14, power=68, all=86) now return in a single page. Eliminates multi-page pagination that caused tool visibility issues in Claude Code and other MCP clients. (#96)

[0.10.1] - 2026-05-18

Added

  • MCP token/context optimizations — 5 optimizations that cut tool description tokens by ~47%: description compression, property deduplication, default value removal, cursor-based pagination, and compact enum formatting. Core profile now ~2,400 estimated tokens (was ~4,500+). (#88)
  • Cross-compilation CI workflow — automated release builds for 10 platform binaries: Linux (x64, ARM64, musl, ARMv7), macOS (ARM64, x64), Windows (x64, ARM64), and FreeBSD (x64). Triggers on GitHub Release publish. (#87)

Fixed

  • Updated documentation to reflect correct tool count: 14 core tools (86+ available via profiles).

[0.10.0] - 2026-05-17

Breaking Changes

  • Tool surface reduced from 169 to 14 default tools — the MCP server now exposes only 14 core memory tools by default, down from 169. Agents discover additional tools via the dakera_discover_tools meta-tool. Saves ~35K tokens per session. (#83)
  • Removed admin, analytics, and streaming tools — dakera_admin_*, dakera_analytics_*, dakera_streaming_*, and redundant graph/vector tools have been permanently deleted (-1,672 lines). If you relied on these tools, use the Dakera REST API directly. (#84)

Added

  • Profile-based tool tiering — three profiles (minimal, standard, full) control which tools are exposed. Set via DAKERA_MCP_PROFILE env var. standard is the default (14 tools). (#83)
  • dakera_discover_tools meta-tool — agents can dynamically discover and load additional tools beyond the default set without exposing them all upfront. (#83)
  • 25 integration tests against live Dakera Docker — end-to-end verification of store, recall, search, session lifecycle, and profile switching against a real Dakera instance. (#85)
  • 47 deep profile tiering validation tests — comprehensive coverage of tool visibility, profile transitions, and meta-tool behavior (190 → 237 total tests). (#86)

Changed

  • Consolidated graph_traverse/graph_path/graph_export into unified graph tools. (#84)
  • Consolidated vector_query/vector_multi_search/vector_explain into unified vector tools. (#84)

Chore

  • Applied ARM rustfmt 1.9.0 formatting.

[0.9.8] - 2026-05-13

Chore

  • Version bump for Docker image rebuild with MCP annotation. (#72)

[0.9.7] - 2026-04-29

Changed

  • Bumped thiserror from 1.0.69 to 2.0.18 — internal error handling improvement, no public API change. (#56)
  • Pinned Rust toolchain to 1.95.0 for reproducible builds. (#59)

Fixed

  • Docker multi-arch release builds — replaced push-by-digest with platform-tagged images (VERSION-amd64, VERSION-arm64) followed by imagetools create manifest merge. Self-hosted ARM runners had unreliable digest availability in GHCR, causing all Docker releases since v0.9.2 to fail.
  • Added release workflow concurrency control to prevent parallel tag pushes from racing.

Chore

  • Updated Cargo.lock patch dependencies. (#60)
  • Added Dependabot configuration for Cargo and GitHub Actions.

[0.9.6] - 2026-04-29

Fixed

  • Concurrent request handling and crash recovery — wraps stdout in Mutex<Stdout> and adds catch_unwind around tool handlers to prevent a single panicking tool from crashing the MCP process. Critical for Claude Code which issues concurrent tool calls over stdio. (#54)

Security

  • Bumped rustls-webpki from 0.103.12 to 0.103.13 (follow-up security patch). (#52)

[0.9.5] - 2026-04-29

Fixed

  • Retry with exponential backoff, connection pooling, and tool timeout — prevents transient server errors from surfacing as MCP tool failures. Adds reqwest connection pooling for keep-alive reuse and a 60-second per-tool timeout. (#53)
  • Nest filter params in batch_recall and batch_forget — filters like tags, min_importance, created_after were sent as top-level fields instead of nested under the correct request body structure. (#51)

CI

  • Removed GHCR digest validation loop from release workflow (DAK-2005). (#50)

[0.9.4] - 2026-04-17

CI

  • Extended GHCR digest retry window and upgraded to Node.js 24 actions. (#45)

Dependencies

  • Bumped rand from 0.9.2 to 0.9.4. (#46)
  • Security — rustls-webpki CVE patch: Updated to rustls-webpki 0.103.12 addressing GHSA-xgp8-3hg3-c2mh and GHSA-965h-392x-2mh5 (CVSS 2.2 LOW). (#48)

[0.9.3] - 2026-04-08

Added

MCP-5 — Cognitive MCP Tools (server v0.9.6+ / v0.9.9+)

Three new tools for managing per-namespace memory policies and Deep Associative Recall directly from MCP clients:

  • dakera_memory_policy_get — read the MemoryPolicy for a namespace. Returns all COG-1, COG-3, and SEC-5 policy fields (differential TTLs, per-type decay curves, spaced-repetition settings, consolidation config, and rate limits). Wraps GET /v1/namespaces/:namespace/memory_policy. Requires Read scope. Server v0.9.6+.
  • dakera_memory_policy_set — partial-update the MemoryPolicy for a namespace. Performs a GET-then-PUT merge internally so omitted fields are preserved; no need to supply the full schema. Wraps PUT /v1/namespaces/:namespace/memory_policy. Requires Write scope. Server v0.9.6+ (SEC-5 rate-limit fields require v0.9.9+).
  • dakera_recall_associated — dedicated Deep Associative Recall (KG-3). Wraps POST /v1/memory/recall with include_associated: true and exposes the KG-3 parameters directly: associated_memories_depth (1–3 hops, default 1) and associated_memories_min_weight (0.0–1.0, default 0.0). Response includes memories (primary hits) and associated_memories (KG neighbours, each with a depth field). Requires Read scope. Server v0.9.8+.

Total MCP tool count: 80 → 83.

Changed

  • Updated README.md to reflect open-core model and current product positioning. (PR#43)
  • Docs: expanded Available Tools reference from 45 → 83 tools with version badge. (PR#41)
  • Docs: filled changelog history gap v0.4.1 → v0.9.1. (PR#40)

CI

  • Release workflow: added GHCR propagation validation step before manifest creation to prevent partial-push releases. (PR#42)

[0.9.1] - 2026-04-01

Changed

  • dakera_recall: added include_associated boolean parameter (default false) — when true, the server traverses the knowledge graph from each primary result and includes neighbouring memories in the response. Backed by COG-2 (server v0.9.6+). (PR#36)
  • dakera_recall: added since and until ISO-8601 string parameters for time-window recall. Both are optional and can be combined with other filters. Backed by CE-7 (server v0.9.7+). (PR#38)

[0.9.0] - 2026-03-31

Added

ODE-2 — dakera-ode Entity Extraction Sidecar

  • dakera_extract_entities — run entity extraction via the dakera-ode sidecar. Calls POST /ode/extract on DAKERA_ODE_URL (default: http://localhost:8080). Accepts content, agent_id, and optional memory_id / entity_types filters. Returns { entities: [{text, label, start, end, score}], model, processing_time_ms }. Requires DAKERA_ODE_URL to be configured. (PR#33)

KG-2 — Knowledge Graph Query & Export (server v0.9.8+, backed by /v1/knowledge/*)

  • dakera_kg_traverse — BFS traversal from a root memory. Accepts edge_type, min_weight, max_depth (default 3), and limit filters. (PR#34)
  • dakera_kg_query — filter-based graph query without requiring a root node; useful for neighbourhood exploration across an agent's full graph. (PR#34)
  • dakera_kg_export — export the knowledge graph as json (structured edges) or graphml (XML compatible with Gephi / Cytoscape / yEd). (PR#34)

Total MCP tool count: 76 → 79 → 80 (KG-2 → v0.9.1 params, no new tools).

[0.8.0] - 2026-03-30

Added

EXT-1 — Pluggable Extraction Providers (server v0.9.x+)

  • dakera_extract — run extraction via the full provider hierarchy (per-request override → namespace default → server default → GLiNER local). Supports gliner, openai, anthropic, openrouter, ollama, none. Wraps POST /v1/extract. (PR#29)
  • dakera_extractor_get — read the namespace-level default extractor configuration. Wraps GET /v1/namespaces/:ns/extractor. (PR#29)
  • dakera_extractor_set — set or clear the namespace default extractor. Wraps PATCH /v1/namespaces/:ns/extractor. (PR#29)

SEC-3 — AES-256-GCM Encryption Key Rotation (server v0.9.x+)

  • dakera_encryption_rotate_key — zero-downtime key rotation; re-encrypts all memories under the new key in a single pass. Accepts an optional namespace filter. Requires SuperAdmin scope. Wraps POST /admin/encryption/rotate-key. (PR#29)

Total MCP tool count: 72 → 76.

[0.7.0] - 2026-03-30

Added

MCP-4 (CE-4) — Entity Extraction Tools (server v0.9.x+)

  • dakera_auto_tag — extract named entities from text via the GLiNER NER pipeline (with a rule-based pre-pass). Wraps POST /v1/memories/extract. (PR#25)
  • dakera_entity_types_set — enable entity extraction and configure entity types per namespace. Wraps PATCH /v1/namespaces/:ns/config. (PR#25)
  • dakera_memory_entities — retrieve the entity tags stored on a memory at write time. Wraps GET /v1/memory/entities/:id. (PR#25)

MCP-4 (CE-5) — Memory Knowledge Graph (server v0.9.x+)

  • dakera_graph_traverse — BFS traversal from a memory node with configurable depth (1–5, default 3). Wraps GET /v1/memories/:id/graph. (PR#26)
  • dakera_graph_path — shortest path between two memories. Wraps GET /v1/memories/:id/path. (PR#26)
  • dakera_graph_link_memory — create an explicit linked_by edge between two memories. Wraps POST /v1/memories/:id/links. (PR#26)
  • dakera_graph_export — export the full knowledge graph for an agent. Wraps GET /v1/agents/:id/graph/export. (PR#26)

DX-1 — Memory Import / Export

  • dakera_memory_export — export agent memories to JSONL, CSV, Mem0, or Zep format. Wraps GET /v1/agents/:id/memories/export. (PR#28)
  • dakera_memory_import — import memories from any supported format (auto-detected). Wraps POST /v1/agents/:id/memories/import. (PR#28)

SEC-1 — Namespace-Scoped API Keys

  • dakera_namespace_key_create — create an API key scoped to a specific namespace. (PR#28)
  • dakera_namespace_key_list — list keys with access to a namespace. (PR#28)
  • dakera_namespace_key_delete — revoke a namespace key. (PR#28)
  • dakera_namespace_key_usage — get usage statistics for a namespace key. (PR#28)

INT-1 — Memory Feedback

  • dakera_memory_feedback — upvote, downvote, or flag a memory for quality signals. (PR#28)
  • dakera_memory_feedback_get — retrieve the feedback history for a specific memory. (PR#28)
  • dakera_agent_feedback_summary — aggregate feedback summary across all memories for an agent. (PR#28)

Total MCP tool count: 54 → 72 (16 CE-4/CE-5 tools from internal v0.5.0/v0.6.0 + DX-1/SEC-1/INT-1).

[0.4.1] - 2026-03-23

Fixed

  • dakera_hybrid_search: vector parameter is now optional, enabling BM25-only full-text (BM25) search without requiring a pre-computed embedding vector. Fixes DAK-679. (PR#20)

[0.4.0] - 2026-03-23

Added

  • dakera_decay_config_get — returns the current memory-decay configuration (strategy, half_life_hours, min_importance). Wraps GET /admin/decay/config (Admin scope). Implements DECAY-1.
  • dakera_decay_config_set — updates decay settings at runtime without a server restart. Accepts optional strategy (exponential | linear | step), half_life_hours (> 0), and min_importance (0.0–1.0). Wraps PUT /admin/decay/config (Admin scope). Implements DECAY-1.
  • dakera_decay_stats — returns cumulative decay counters (total_decayed, total_deleted, cycles_run, last_run_at) plus per-cycle detail for the most recent run. Wraps GET /admin/decay/stats (Admin scope). Implements DECAY-2.
  • dakera_store schema updated — new optional expires_at parameter (Unix timestamp in seconds). Takes precedence over ttl_seconds; on expiry the memory is hard-deleted by the decay engine.

Tests

  • Unit tests for dakera_decay_config_set covering invalid strategy, zero/negative half_life_hours, and out-of-range min_importance; coverage for definitions length, non-empty descriptions, and unknown tool dispatch for the decay module.

[0.3.2] - 2026-03-23

Added

  • dakera_autopilot_status — returns the live AutoPilot configuration plus last-run statistics (timestamps, memories deduped, clusters consolidated). Wraps GET /admin/autopilot/status (Admin scope). Implements PILOT-4.
  • dakera_autopilot_trigger — forces an immediate AutoPilot cycle. Accepts action of dedup, consolidate, or all. Wraps POST /admin/autopilot/trigger (Admin scope). Implements PILOT-4.

Tests

  • Unit tests for dakera_autopilot_trigger covering invalid action, missing action, and unknown tool dispatch; and for dakera_autopilot_status / dakera_autopilot_trigger definitions.

[0.3.1] - 2026-03-22

Added

  • Native binary release artifact: CI now builds and uploads a musl-linked dakera-mcp binary on every GitHub release (ci: deploy native binary on release, DAK-575).

Tests

  • Unit tests for dakera_batch_recall and dakera_batch_forget MCP tools, covering request serialization and response deserialization (test: unit tests for batch recall/forget, v0.3.0 gap).

Security

  • Pinned all appleboy GitHub Actions to exact SHA digests to harden CI against supply-chain attacks (fix(ci): pin appleboy actions to exact versions).

[0.3.0] - 2026-03-22

Added

  • dakera_batch_recall — filter-based bulk recall without semantic search (CE-2, dakera server v0.7.0). Accepts tags, min_importance, max_importance, created_after, created_before, memory_type, and session_id predicates. Calls POST /v1/memories/recall/batch.
  • dakera_batch_forget — filter-based bulk delete (CE-2, dakera server v0.7.0). Same predicate set as dakera_batch_recall. The server enforces at least one filter to prevent accidental full-namespace wipe. Calls DELETE /v1/memories/forget/batch with a JSON body.
  • delete_with_json internal API client method to support DELETE requests carrying a JSON body.

[0.2.2] - 2026-03-21

Added

  • dakera_namespace_configure — create-or-update (upsert) a namespace via PUT /v1/namespaces/:namespace. Added in dakera server v0.6.0. Requires Write scope.
  • dakera_knowledge_network_cross_agent — build a cross-agent memory similarity network via POST /v1/knowledge/network/cross-agent. Added in dakera server v0.4.0 (DASH-A). Requires Admin scope.

[0.2.1] - 2026-03-20

Fixed

  • Corrected execute_tool placement (moved before test module) and rustfmt formatting

Added

  • Unit tests for protocol types and tool helpers (DAK-173)

Security

  • Add explicit GITHUB_TOKEN permissions to CI workflow (#1)

Chore

  • Upgrade GitHub Actions runners to Node.js 24 compatible versions

[0.2.0] - 2025-03-15

Added

  • MCP server with stdio JSON-RPC transport
  • Memory tools: store, recall, search, get, update, importance, forget, consolidate
  • Session tools: start, end, list, get, memories
  • Agent tools: stats, memories, sessions
  • Knowledge tools: graph, summarize, deduplicate
  • Namespace tools: list, get, create, delete
  • Vector tools: upsert, query, delete, batch query, bulk update, bulk delete, count, export, aggregate, multi-search, upsert columns, explain, warm, unified query
  • Full-text search tools: index, search, delete, stats, hybrid search
  • Inference tools: text query, upsert text, batch query text
  • Docker support with multi-stage build
  • Bearer token authentication support